Abstract Purpose Scientists and producers have pursued precision agriculture to increase yields and limit environmental impacts. Applying precision agriculture concepts has not been straightforward, however, because complex interactions between soil and weather govern crop growth. We hypothesize that evapotranspiration (ET) is a single metric that captures interactions between plant status, soil, topography, and weather and can be used to predict spatial patterns in crop yield. Methods We used remotely sensed estimates of ET to calculate a normalized ET metric describing the ratio of actual to reference ET (f RET ) at a research farm located in the U.S. Corn Belt. Using the resulting 30-m resolution maps of f RET , we estimated the within-field variability in crop water stress and used machine learning techniques to relate f RET , soil, and topographic properties to crop yield. Results We show that total growing season ET is not a strong predictor of yield spatial variability ( r < 0.3), but when used to train a random forest model, f RET estimates are stronger predictors (mean R 2 = 0.6). In fact, f RET alone is better at predicting yield than any other combination of soil, topographic, and hydrologic data tested with machine learning methods. We also show that end of season yield can be predicted with just 9 weeks of f RET data which provides the opportunity to use remotely sensed ET to guide in-season site-specific management decisions. Conclusion This novel combination of earth observations and machine learning algorithms can guide sustainable intensification of agricultural systems within the context of increasing water stress.
Schreiner-McGraw et al. (Fri,) studied this question.